CN117874619A - Customer service scoring model construction method and device, computer equipment and storage medium - Google Patents

Abstract

The application relates to a customer service scoring model construction method, a customer service scoring model construction device, computer equipment and a storage medium. The application relates to the technical field of artificial intelligence and deep learning. The method comprises the following steps: acquiring a plurality of unlabeled customer service data, real label information thereof and a plurality of identification teacher networks, and identifying each customer service data and data types of each identification teacher network; based on the recognition teacher network, predicting pseudo tag information of each customer service data, and screening target customer service data corresponding to noise values lower than a noise threshold; generating a transition probability matrix of the data type based on transition probability information between pseudo tag information and real tag information of each target customer service data; training an initial recognition student model through a knowledge distillation loss function based on the information to obtain a recognition student model corresponding to the data type; and obtaining a scoring strategy of each data type, thereby constructing a customer service scoring model. By adopting the method, the training efficiency of the customer service scoring model can be improved.

Description

Customer service scoring model construction method and device, computer equipment and storage medium

Technical Field

The application relates to the technical field of artificial intelligence and deep learning, in particular to a customer service scoring model construction method, a customer service scoring model construction device, computer equipment and a storage medium.

Background

Along with the continuous development of customer service, in order to better identify customer service quality, banks often score customer service to identify defects and defects of the customer service, and timely iterate and update the customer service to improve the customer service quality, but often score the customer service to consume time and experience of the customer, thereby influencing the customer to evaluate the customer service, so how to self-evaluate the customer service quality is a current research problem.

The traditional self-evaluation mode of customer service quality is to construct a model through traditional OpenPose and HMM, and the final service score of customer service is obtained by fusing body gesture recognition results and voice keyword recognition results in decision. However, by using the traditional model construction method of OpenPose and HMM, the generated neural network consists of a large number of parameters, so that the resource consumption is high, the calculation cost is high, and the neural network cannot be used in low-performance small-sized equipment; meanwhile, when the light model is trained by using original training data (such as private data, human body video, photo and voice) of the large model, the problems of data safety and privacy protection can be generated, so that the training efficiency of the customer service scoring model is lower.

Disclosure of Invention

In view of the foregoing, it is desirable to provide a method, apparatus, computer device, computer-readable storage medium, and computer program product for constructing a customer service scoring model.

In a first aspect, the present application provides a method for constructing a customer service scoring model. The method comprises the following steps:

acquiring a plurality of unlabeled customer service data, real label information of each customer service data and a plurality of trained identification teacher networks, and identifying the data type of each customer service data and the data type corresponding to each identification teacher network;

for each data type, based on an identification teacher network of the data type, respectively carrying out data marking processing on each customer service data of the data type to obtain pseudo tag information of each customer service data, and calculating a noise value of each customer service data based on the pseudo tag information of each customer service data and the real tag information of each customer service data;

screening customer service data corresponding to a noise value lower than a noise threshold value, serving as target customer service data of the data type, and generating a transition probability matrix of the data type based on pseudo tag information of each target customer service data and transition probability information between real tag information of each target customer service data;

Training classification parameters of an initial recognition student model corresponding to the data type through a knowledge distillation loss function based on pseudo tag information of each target customer service data, a conversion probability matrix of the data type, a recognition teacher network of the data type and each target customer service data to obtain the recognition student model corresponding to the data type;

and obtaining a scoring strategy of each data type, and constructing a customer service scoring model based on the identification model corresponding to each data type and the scoring strategy of each data type.

Optionally, the identifying the data type of each customer service data and the data type corresponding to each identified teacher network includes:

extracting data identification information of each customer service data, and identifying the data type of each customer service data based on the data identification information of each customer service data;

acquiring training sample data of each recognition teacher network and data identification information of each training sample data, and recognizing the data type of each training sample data based on the data identification information of each training sample data; the training process of the teacher network is that the local client performs parameter training processing on each initial teacher network based on training sample data of each teacher network to obtain each teacher network;

And taking the data type of each training sample data as the data type of the recognition teacher network trained by each training sample data.

Optionally, the identifying teacher network based on the data type performs data marking processing on each customer service data of the data type to obtain pseudo tag information of each customer service data, and calculates a noise value of each customer service data based on the pseudo tag information of each customer service data and the real tag information of each customer service data, including:

based on the data type recognition teacher network, predicting the predicted data category of each customer service data respectively, and taking the predicted data category of each customer service data as pseudo tag information of each customer service data;

and respectively calculating the noise value of each customer service data through a label-free data noise value algorithm based on each customer service data and the pseudo label information of each customer service data.

Optionally, the generating the transition probability matrix of the data type based on the pseudo tag information of each target customer service data and the transition probability information between the real tag information of each target customer service data includes:

Identifying the data category of the pseudo tag information of each target customer service data and the data category corresponding to the real tag information of each target customer service data;

based on the target customer service data to which each piece of pseudo tag information belongs, the target customer service data to which each piece of real tag information belongs, the data category corresponding to each piece of pseudo tag information, and the data category corresponding to each piece of real tag information, respectively calculating the conversion probability of each piece of real tag information into the pseudo tag information through a probability calculation formula;

and constructing a transition probability matrix of the data type based on the transition probability of each real tag information to a pseudo tag information, each pseudo tag information and each real tag information.

Optionally, the training, by using a knowledge distillation loss function, classification parameters of the initial recognition student model corresponding to the data type to obtain the recognition student model corresponding to the data type based on the pseudo tag information of each target customer service data, the transition probability matrix of the data type, the recognition teacher network of the data type, and each target customer service data, includes:

Based on the transition probability matrix of the data type, adjusting an initial knowledge distillation loss function to obtain a knowledge distillation loss function corresponding to the data type;

based on the pseudo tag information of each target customer service data, the identification teacher network of the data type and each target customer service data, training the classification parameters of the initial identification student model corresponding to the data type through the knowledge distillation loss function of the data type to obtain the identification student model corresponding to the data type.

Optionally, the building a customer service scoring model based on the identification model corresponding to each data type and the scoring policy of each data type includes:

the method comprises the steps of obtaining grading information of real tag information of each data type and grading weight values of each data type, and respectively adding the grading information of the real tag information of each data type to an identification student model corresponding to each data type to obtain customer service grading student models corresponding to each data type;

and determining weight parameters of customer service scoring student models corresponding to the data types based on the scoring weight values of the data types, and summarizing the customer service scoring student models and the weight parameters of the customer service scoring student models to obtain the customer service scoring models.

In a second aspect, the application further provides a device for constructing the customer service scoring model. The device comprises:

the system comprises an acquisition module, a processing module and a processing module, wherein the acquisition module is used for acquiring a plurality of unlabeled customer service data, real label information of each customer service data and a plurality of trained recognition teacher networks, and recognizing the data type of each customer service data and the data type corresponding to each recognition teacher network;

the computing module is used for respectively carrying out data marking processing on each customer service data of the data type according to the identification teacher network of the data type to obtain pseudo tag information of each customer service data, and computing the noise value of each customer service data according to the pseudo tag information of each customer service data and the real tag information of each customer service data;

the screening module is used for screening customer service data corresponding to a noise value lower than a noise threshold value, serving as target customer service data of the data type, and generating a transition probability matrix of the data type based on pseudo tag information of each target customer service data and transition probability information between the pseudo tag information and real tag information of each target customer service data;

The training module is used for training the classification parameters of the initial recognition student model corresponding to the data type through a knowledge distillation loss function based on the pseudo tag information of each target customer service data, the conversion probability matrix of the data type, the recognition teacher network of the data type and each target customer service data to obtain the recognition student model corresponding to the data type;

the construction module is used for acquiring the scoring strategies of the data types and constructing a customer service scoring model based on the identification model corresponding to the data types and the scoring strategies of the data types.

Optionally, the acquiring module is specifically configured to:

extracting data identification information of each customer service data, and identifying the data type of each customer service data based on the data identification information of each customer service data;

acquiring training sample data of each recognition teacher network and data identification information of each training sample data, and recognizing the data type of each training sample data based on the data identification information of each training sample data; the training process of the teacher network is that the local client performs parameter training processing on each initial teacher network based on training sample data of each teacher network to obtain each teacher network;

And taking the data type of each training sample data as the data type of the recognition teacher network trained by each training sample data.

Optionally, the computing module is specifically configured to:

based on the data type recognition teacher network, predicting the predicted data category of each customer service data respectively, and taking the predicted data category of each customer service data as pseudo tag information of each customer service data;

and respectively calculating the noise value of each customer service data through a label-free data noise value algorithm based on each customer service data and the pseudo label information of each customer service data.

Optionally, the screening module is specifically configured to:

identifying the data category of the pseudo tag information of each target customer service data and the data category corresponding to the real tag information of each target customer service data;

based on the target customer service data to which each piece of pseudo tag information belongs, the target customer service data to which each piece of real tag information belongs, the data category corresponding to each piece of pseudo tag information, and the data category corresponding to each piece of real tag information, respectively calculating the conversion probability of each piece of real tag information into the pseudo tag information through a probability calculation formula;

And constructing a transition probability matrix of the data type based on the transition probability of each real tag information to a pseudo tag information, each pseudo tag information and each real tag information.

Optionally, the training module is specifically configured to:

based on the transition probability matrix of the data type, adjusting an initial knowledge distillation loss function to obtain a knowledge distillation loss function corresponding to the data type;

based on the pseudo tag information of each target customer service data, the identification teacher network of the data type and each target customer service data, training the classification parameters of the initial identification student model corresponding to the data type through the knowledge distillation loss function of the data type to obtain the identification student model corresponding to the data type.

Optionally, the construction module is specifically configured to:

the method comprises the steps of obtaining grading information of real tag information of each data type and grading weight values of each data type, and respectively adding the grading information of the real tag information of each data type to an identification student model corresponding to each data type to obtain customer service grading student models corresponding to each data type;

And determining weight parameters of customer service scoring student models corresponding to the data types based on the scoring weight values of the data types, and summarizing the customer service scoring student models and the weight parameters of the customer service scoring student models to obtain the customer service scoring models.

In a third aspect, the present application provides a computer device. The computer device comprises a memory storing a computer program and a processor implementing the steps of the method of any of the first aspects when the processor executes the computer program.

In a fourth aspect, the present application provides a computer-readable storage medium. On which a computer program is stored which, when being executed by a processor, implements the steps of the method of any of the first aspects.

In a fifth aspect, the present application provides a computer program product. The computer program product comprising a computer program which, when executed by a processor, implements the steps of the method of any of the first aspects.

The construction method, the construction device, the computer equipment and the storage medium of the customer service scoring model are characterized in that a plurality of unlabeled customer service data and a plurality of trained identification teacher networks are obtained, and the data type of each customer service data and the data type corresponding to each identification teacher network are identified; for each data type, based on an identification teacher network of the data type, respectively carrying out data marking processing on each customer service data of the data type to obtain pseudo tag information of each customer service data, and based on the pseudo tag information of each customer service data, calculating a noise value of each customer service data; screening customer service data corresponding to a noise value lower than a noise threshold value, serving as target customer service data of the data type, and generating a transition probability matrix of the data type based on pseudo tag information of each target customer service data and transition probability information between real tag information of each customer service data; training classification parameters of an initial recognition student model corresponding to the data type through a knowledge distillation loss function based on pseudo tag information of each target customer service data, a conversion probability matrix of the data type and each target customer service data to obtain the recognition student model corresponding to the data type; and obtaining a scoring strategy of each data type, and constructing a customer service scoring model based on the identification model corresponding to each data type and the scoring strategy of each data type. According to the scheme, the classification parameters of the initial student model of each data type are trained based on the trained identification teacher network of each data type through the knowledge distillation strategy, so that the problems that a large number of parameters are required for constructing the model, the resource consumption is overlarge and the calculation cost is high are avoided, a simplified identification model of each data type is constructed, and the resource space and the calculation capacity of detection equipment are saved. Secondly, each recognition teacher network is a recognition network which is trained on a local client, so that the problem that private data is revealed by each recognition teacher network of client data is avoided, and the data safety and privacy protection effect in the process of training the recognition network are improved. And then, through calculating the conversion probability information of the pseudo tag information and the real tag information, after generating the conversion probability matrix of each data type, the classification parameters of each initial recognition student model are trained, so that the problem of data noise of the pseudo tag information predicted by the recognition teacher network is avoided, and the recognition accuracy of the initial student recognition network trained by the knowledge distillation strategy based on the recognition teacher network is improved. And finally, training an initial student identification network and a scoring strategy of each data type through the unlabeled customer service data and the trained identification teacher network to construct a customer service scoring mode, so that the sample data quantity participating in training is reduced, and the training efficiency of the customer service scoring model is improved under the condition that the training accuracy is ensured.

Drawings

FIG. 1 is a flow diagram of a method of constructing a customer scoring model in one embodiment;

FIG. 2 is a flow diagram of an example of the construction of a customer scoring model in one embodiment;

FIG. 3 is a block diagram of an apparatus for constructing a customer scoring model in one embodiment;

fig. 4 is an internal structural diagram of a computer device in one embodiment.

Detailed Description

In order to make the objects, technical solutions and advantages of the present application more apparent, the present application will be further described in detail with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are for purposes of illustration only and are not intended to limit the present application.

The customer service scoring model construction method provided by the embodiment of the application can be applied to an application environment for customer service scoring model construction. The method can be applied to the terminal, the server and a system comprising the terminal and the server, and is realized through interaction of the terminal and the server. The terminal may be, but not limited to, various personal computers, notebook computers, smart phones, tablet computers, and the like. The terminal trains the classification parameters of the initial student model of each data type based on the trained identification teacher network of each data type through a knowledge distillation strategy, so that the problems that a large number of parameters are needed for constructing the model, the resource consumption is overlarge and the calculation cost is high are avoided, a simplified identification model of each data type is constructed, and the resource space and the calculation capacity of detection equipment are saved. Secondly, each recognition teacher network is a recognition network which is trained on a local client, so that the problem that private data is revealed by each recognition teacher network of client data is avoided, and the data safety and privacy protection effect in the process of training the recognition network are improved. And then, through calculating the conversion probability information of the pseudo tag information and the real tag information, after generating the conversion probability matrix of each data type, the classification parameters of each initial recognition student model are trained, so that the problem of data noise of the pseudo tag information predicted by the recognition teacher network is avoided, and the recognition accuracy of the initial student recognition network trained by the knowledge distillation strategy based on the recognition teacher network is improved. And finally, training an initial student identification network and a scoring strategy of each data type through the unlabeled customer service data and the trained identification teacher network to construct a customer service scoring mode, so that the sample data quantity participating in training is reduced, and the training efficiency of the customer service scoring model is improved under the condition that the training accuracy is ensured.

In one embodiment, as shown in fig. 1, a method for constructing a customer service scoring model is provided, and the method is applied to a terminal for illustration, and includes the following steps:

step S101, acquiring a plurality of unlabeled customer service data, real label information of each customer service data and a plurality of trained teacher identification networks, and identifying the data type of each customer service data and the data type corresponding to each teacher identification network.

In this embodiment, the terminal obtains historical customer service data of different data types in the historical customer service database, and removes data tag information of each historical customer service data to obtain a plurality of unlabeled customer service data. Wherein the data type is a customer service state type characterized by the data, and the customer service state type comprises, but is not limited to, a body posture type, a voice word type and the like. The data tag information of each customer service data is a data category corresponding to the customer service data, for example, the data tag information of the body posture type comprises sitting, chin rest, hand lifting, low head, askew sitting and the like. Then, the terminal trains each initial recognition teacher network through a plurality of local sample data with different data types, and obtains the recognition teacher network with each data type. The training mode of each initial teacher network is that based on training sample data (namely local sample data) of each identification teacher network, the local client performs classification parameter training processing on each initial identification teacher network to obtain each identification teacher network. Then, the terminal takes the data tag information of each customer service data as the real tag information of each customer service data. And finally, the terminal identifies the data type of each customer service data and the data type corresponding to each identified teacher network. The specific data type identification process will be described in detail later. The data type corresponding to the teacher identification network is the data type of training sample data for training the teacher identification network. The teacher neural network corresponding to the knowledge distillation technology is identified as the teacher network. And identifying the student network as a student neural network corresponding to the knowledge distillation technology. The teacher network for recognizing the body posture type recognizes the convolutional neural network for image classification.

Step S102, for each data type, based on the identification teacher network of the data type, performing data marking processing on each customer service data of the data type to obtain pseudo tag information of each customer service data, and calculating the noise value of each customer service data based on the pseudo tag information of each customer service data.

In this embodiment, the terminal performs data marking processing on each customer service data of a data type according to the identification teacher network of the data type to obtain pseudo tag information of each customer service data, and calculates a noise value of each customer service data based on the pseudo tag information of each customer service data. The pseudo tag information is data tag information based on customer service data predicted by the recognition classroom network. The noise value of the customer service data is used for representing the classification deviation value between the pseudo tag information of each customer service data and the real tag data of each customer service data. Wherein the smaller the noise value, the greater the availability of the customer service data is characterized.

Step S103, screening customer service data corresponding to noise values lower than a noise threshold value, taking the customer service data as target customer service data of a data type, and generating a transition probability matrix of the data type based on pseudo tag information of each target customer service data and transition probability information between the pseudo tag information of each target customer service data and real tag information of each target customer service data.

In this embodiment, a noise threshold is preset in a terminal, customer service data corresponding to a noise value lower than the noise threshold is screened and used as target customer service data of a data type, and a transition probability matrix of the data type is generated based on pseudo tag information of each target customer service data and transition probability information between the pseudo tag information and real tag information of each target customer service data. The specific calculation process of the transition probability information and the generation process of the transition probability matrix will be described in detail later. The conversion probability information predicts the real tag information of the target customer service data as the conversion probability of the pseudo tag information when the recognition teacher network of each data type predicts each target customer service data, namely the conversion probability characterizes the noise probability of the pseudo tag information of the recognition teacher network of the data type.

Step S104, training classification parameters of the initial recognition student model corresponding to the data type through a knowledge distillation loss function based on pseudo tag information of each target customer service data, a conversion probability matrix of the data type, a recognition teacher network of the data type and each target customer service data, and obtaining the recognition student model corresponding to the data type.

In this embodiment, the terminal trains classification parameters of the initial recognition student model corresponding to the data type through the knowledge distillation loss function based on pseudo tag information of each target customer service data, a transition probability matrix of the data type, a recognition teacher network of the data type, and each target customer service data, and obtains the recognition student model corresponding to the data type. The specific training process will be described in detail later. The knowledge distillation loss function is obtained by adding a conversion probability matrix of each data type to the conventional knowledge distillation loss function.

Step S105, a scoring strategy of each data type is obtained, and a customer service scoring model is constructed based on the identification model corresponding to each data type and the scoring strategy of each data type.

In this embodiment, the terminal obtains a scoring policy of each data type, and constructs a customer service scoring model based on the identification model corresponding to each data type and the scoring policy of each data type. The scoring strategies of the data types are stored in the databases and used for scoring the recognition results of the customer service data of the data types by the recognition student network of the data types.

Based on the scheme, the classification parameters of the initial student model of each data type are trained based on the trained identification teacher network of each data type through the knowledge distillation strategy, so that the problems that a large number of parameters are required for constructing the model, the resource consumption is overlarge and the calculation cost is high are avoided, a simplified identification model of each data type is constructed, and the resource space and the calculation capacity of the detection equipment are saved. Secondly, each recognition teacher network is a recognition network which is trained on a local client, so that the problem that private data is revealed by each recognition teacher network of client data is avoided, and the data safety and privacy protection effect in the process of training the recognition network are improved. And then, through calculating the conversion probability information of the pseudo tag information and the real tag information, after generating the conversion probability matrix of each data type, the classification parameters of each initial recognition student model are trained, so that the problem of data noise of the pseudo tag information predicted by the recognition teacher network is avoided, and the recognition accuracy of the initial student recognition network trained by the knowledge distillation strategy based on the recognition teacher network is improved. And finally, training an initial student identification network and a scoring strategy of each data type through the unlabeled customer service data and the trained identification teacher network to construct a customer service scoring mode, so that the sample data quantity participating in training is reduced, and the training efficiency of the customer service scoring model is improved under the condition that the training accuracy is ensured.

Optionally, identifying the data type of each customer service data and the data type corresponding to each identified teacher network includes: extracting data identification information of each customer service data, and identifying the data type of each customer service data based on the data identification information of each customer service data; acquiring training sample data of each recognition teacher network and data identification information of each training sample data, and recognizing the data type of each training sample data based on the data identification information of each training sample data; the training process of the teacher network is that the local client performs parameter training processing on each initial teacher network based on training sample data of each teacher network to obtain each teacher network; and taking the data type of each training sample data as the data type of the recognition teacher network trained by each training sample data.

In this embodiment, the terminal extracts the data identification information of each customer service data, and identifies the data type of each customer service data based on the data identification information of each customer service data. The data identification information is identification information which is contained in each customer service data and used for representing the data type of the customer service data. The data identification information is the data name of the customer service data. The different data types comprise different data ranges, the terminal queries the data name range to which each data name belongs, determines the number type corresponding to each data name, for example, the data name of customer service data of the body posture type comprises front image data, side image data, top image data and the like

The terminal acquires training sample data of each recognition teacher network and data identification information of each training sample data, and recognizes the data type of each training sample data based on the data identification information of each training sample data. The training process of the teacher network is that the local client performs parameter training processing on each initial teacher network based on training sample data of each teacher network to obtain each teacher network.

And finally, the terminal takes the data type of each training sample data as the data type of the recognition teacher network trained by each training sample data.

Based on the scheme, through identifying the data identification information, each customer service data and the data type of the teacher network are determined, and the data type identification efficiency is improved.

Optionally, based on the data type identification teacher network, performing data marking processing on each customer service data of the data type to obtain pseudo tag information of each customer service data, and calculating a noise value of each customer service data based on the pseudo tag information of each customer service data, including: based on the recognition teacher network of the data type, predicting the predicted data category of each customer service data respectively, and taking the predicted data category of each customer service data as pseudo tag information of each customer service data; and respectively calculating the noise value of each customer service data through a label-free data noise value algorithm based on each customer service data and the pseudo label information of each customer service data.

In this embodiment, the terminal predicts the predicted data type of each customer service data based on the data type identification teacher network, and uses the predicted data type of each customer service data as pseudo tag information of each customer service data.

The calculation formula for predicting the pseudo tag information of each customer service data is as follows:

in the above, N ^T To identify teacher network, x ^U ∈X ^U For each unlabeled customer service data, yi is real label information of the ith customer service data, and i is a virtual number of each customer service data.

The terminal calculates the noise value of each customer service data through a label-free data noise value algorithm based on the customer service data and the pseudo label information of the customer service data.

The calculation formula of the calculated noise value is as follows:

in the above, N ^T In order to identify the teacher's network,x predicted for identifying teacher network ^U Pseudo tag information of x ^U ∈X ^U For each unlabeled customer service data, D _KL Is KL divergence (relative entropy).

Based on the scheme, the noise value of each customer service data is calculated by predicting the pseudo tag information of each customer service data, and because each recognition teacher network is a teacher network trained by a large number of sample customer service data, the error probability of the sample information with higher confidence is very low, so that the noise value of each customer service data is calculated based on the recognition teacher network, and the accuracy of calculating the noise value is improved.

Optionally, generating the transition probability matrix of the data type based on the pseudo tag information of each target customer service data and the transition probability information between the real tag information of each target customer service data includes: identifying the data category of the pseudo tag information of each target customer service data and the data category corresponding to the real tag information of each target customer service data; based on the target customer service data to which each piece of pseudo tag information belongs, the target customer service data to which each piece of real tag information belongs, the data category corresponding to each piece of pseudo tag information, and the data category corresponding to each piece of real tag information, respectively calculating the conversion probability of each piece of real tag information into the pseudo tag information through a probability calculation formula; and constructing a transition probability matrix of the data type based on the transition probability of each real tag information to the pseudo tag information, each pseudo tag information and each real tag information.

In this embodiment, the terminal identifies the data category of the pseudo tag information of each target customer service data and the data category corresponding to the real tag information of each target customer service data. The data category is a category corresponding to the data tag information of the target customer service data. The categories of different data tag information are different, and each data tag information corresponds to one data category. For example, the data category corresponding to the sitting data tag information is a sitting category, the data category corresponding to the chin rest data tag information is a chin rest category, the data category corresponding to the hand lifting data tag information is a hand lifting category, the data category corresponding to the low head data tag information is a low head category, and the data category corresponding to the askew sitting data tag information is a askew sitting category.

The terminal calculates the conversion probability of converting each real tag information into the pseudo tag information through a probability calculation formula based on the target customer service data to which each pseudo tag information belongs, the target customer service data to which each real tag information belongs, the data category corresponding to each pseudo tag information, and the data category corresponding to each real tag information.

Specifically, the probability calculation formula is:

in the above formula, k is the data class number of y,the virtual label information is pseudo label information, y is real label information, x is target customer service data, j is virtual number of the real label information of the target customer service data, and i is virtual number of the pseudo label information of the target customer service data.

And finally, the terminal constructs a conversion probability matrix of the data type based on the conversion probability of converting each real tag information into the pseudo tag information, each pseudo tag information and each real tag information.

Specifically, the calculation formula of the transition probability matrix is:

is the transition probability matrix Q, p (y|x) is the transition probability of the real tag,/>Is the transition probability of the pseudo tag information.

Based on the scheme, the training accuracy of training the initial identification student model is improved by identifying the real label information and the conversion probability of the label information.

Optionally, based on the pseudo tag information of each target customer service data, the conversion probability matrix of the data type, the recognition teacher network of the data type, and each target customer service data, training the classification parameters of the initial recognition student model corresponding to the data type through the knowledge distillation loss function to obtain the recognition student model corresponding to the data type, including: based on a transition probability matrix of the data type, adjusting an initial knowledge distillation loss function to obtain a knowledge distillation loss function corresponding to the data type; based on the pseudo tag information of each target customer service data, the recognition teacher network of the data type and each target customer service data, training the classification parameters of the initial recognition student model corresponding to the data type through the knowledge distillation loss function of the data type to obtain the recognition student model corresponding to the data type.

In this embodiment, the terminal adjusts the initial knowledge distillation loss function based on the transition probability matrix of the data type, and obtains the knowledge distillation loss function corresponding to the data type.

The calculation formula of the initial knowledge distillation loss function is as follows:

in the above formula, y is real tag information, L ^(hard) Is the cross entropy loss and λ is the weight parameter used to keep the balance of true loss and distillation loss. D (D) _KL As a divergence value, the initial student network learns directly to the recognition teacher network using unlabeled customer service data to obtain the divergence value. N (N) ^T In order to identify the teacher's network,and the target customer service data is x-target customer service data, which is pseudo tag information.

Based on the pseudo tag information of each target customer service data, the recognition teacher network of the data type and each target customer service data, training the classification parameters of the initial recognition student model corresponding to the data type through the knowledge distillation loss function of the data type to obtain the recognition student model corresponding to the data type.

The calculation formula of the knowledge distillation loss function of each data type is as follows:

in the above formula, y is real tag information, L ^(hard) Is the cross entropy loss and λ is the weight parameter used to keep the balance of true loss and distillation loss. D (D) _KL As a divergence value, the initial student network learns directly to the recognition teacher network using unlabeled customer service data to obtain the divergence value. N (N) ^T In order to identify the teacher's network,and (3) taking the pseudo tag information, x target customer service data, and Q as a conversion probability matrix of the data type.

Based on the scheme, the knowledge distillation loss functions are adjusted through the conversion probability matrix of each data type, so that the training effect of training the initial recognition student model of the data type based on the knowledge distillation loss functions is improved, and the recognition accuracy of the recognition student model is improved.

Optionally, constructing a customer service scoring model based on the identification model corresponding to each data type and the scoring policy of each data type includes: the method comprises the steps of obtaining grading information of real label information of each data type and grading weight value of each data type, and respectively adding the grading information of the real label information of each data type to an identification student model corresponding to each data type to obtain customer service grading student models corresponding to each data type; and determining weight parameters of customer service scoring student models corresponding to the data types based on the scoring weight values of the data types, and summarizing the customer service scoring student models and the weight parameters of the customer service scoring student models to obtain the customer service scoring models.

In this embodiment, the terminal obtains the scoring information of each real tag information of each data type and the scoring weight value of each data type, and adds the scoring information of each real tag information of each data type to the identified student model corresponding to each data type, so as to obtain the customer service scoring student model corresponding to each data type. The scoring weight value of each data type is obtained by summarizing and analyzing a large amount of working experience, expert opinion and Internet big data of a worker.

And finally, the terminal determines the weight parameters of the customer service scoring student models corresponding to the data types based on the scoring weight values of the data types, and performs summarization processing on the customer service scoring student models and the weight parameters of the customer service scoring student models to obtain the customer service scoring models.

Based on the scheme, the customer service scoring mode is constructed by training the initial student identification network and the scoring strategy of each data type through the label-free customer service data and the trained identification teacher network, so that the sample data quantity participating in training is reduced, and the training efficiency of the customer service scoring model is improved under the condition that the training accuracy is ensured.

The application also provides a construction example of the customer service scoring model, as shown in fig. 2, and the specific processing process comprises the following steps:

step S201, acquiring a plurality of unlabeled customer service data, real label information of each customer service data, and a plurality of trained recognition teacher networks.

Step S202, extracting data identification information of each customer service data, and identifying the data type of each customer service data based on the data identification information of each customer service data.

Step S203, acquiring training sample data of each teacher identification network and data identification information of each training sample data, and identifying data types of each training sample data based on the data identification information of each training sample data.

Step S204, the data type of each training sample data is used as the data type of the recognition teacher network trained by each training sample data.

Step S205, for each data type, based on the data type identification teacher network, predicting the predicted data type of each customer service data, and using the predicted data type of each customer service data as the pseudo tag information of each customer service data.

Step S206, calculating the noise value of each customer service data through a label-free data noise value algorithm based on each customer service data and the pseudo label information of each customer service data.

Step S207, screening customer service data corresponding to noise values lower than the noise threshold as target customer service data of the data type.

Step S208, identifying the data category of the pseudo tag information of each target customer service data and the data category corresponding to the real tag information of each target customer service data.

Step S209, based on the target customer service data to which each piece of pseudo tag information belongs, the target customer service data to which each piece of real tag information belongs, the data category corresponding to each piece of pseudo tag information, and the data category corresponding to each piece of real tag information, the conversion probability of each piece of real tag information into the pseudo tag information is calculated by a probability calculation formula.

Step S210, a conversion probability matrix of the data type is constructed based on the conversion probability of each real label information into the pseudo label information, each pseudo label information, and each real label information.

Step S211, based on the conversion probability matrix of the data type, the initial knowledge distillation loss function is adjusted, and the knowledge distillation loss function corresponding to the data type is obtained.

Step S212, training classification parameters of the initial recognition student model corresponding to the data type through a knowledge distillation loss function of the data type based on the pseudo tag information of each target customer service data, the recognition teacher network of the data type and each target customer service data, and obtaining the recognition student model corresponding to the data type.

Step S213, obtaining the grading information of each real label information of each data type and the grading weight value of each data type, and respectively adding the grading information of each real label information of each data type to the identification student model corresponding to each data type to obtain the customer service grading student model corresponding to each data type.

Step S214, based on the grading weight value of each data type, determining the weight parameter of the customer service grading student model corresponding to each data type, and summarizing each customer service grading student model and the weight parameter of each customer service grading student model to obtain the customer service grading model.

It should be understood that, although the steps in the flowcharts related to the embodiments described above are sequentially shown as indicated by arrows, these steps are not necessarily sequentially performed in the order indicated by the arrows. The steps are not strictly limited to the order of execution unless explicitly recited herein, and the steps may be executed in other orders. Moreover, at least some of the steps in the flowcharts described in the above embodiments may include a plurality of steps or a plurality of stages, which are not necessarily performed at the same time, but may be performed at different times, and the order of the steps or stages is not necessarily performed sequentially, but may be performed alternately or alternately with at least some of the other steps or stages.

Based on the same inventive concept, the embodiment of the application also provides a construction device of the customer service scoring model for realizing the construction method of the customer service scoring model. The implementation scheme of the solution provided by the device is similar to the implementation scheme recorded in the method, so the specific limitation in the embodiment of the device for constructing one or more customer service scoring models provided below can be referred to the limitation of the method for constructing the customer service scoring model hereinabove, and will not be repeated here.

In one embodiment, as shown in fig. 3, there is provided a device for constructing a customer service scoring model, including: an acquisition module 310, a calculation module 320, a screening module 330, a training module 340, and a construction module 350, wherein:

an obtaining module 310, configured to obtain a plurality of unlabeled customer service data, real label information of each customer service data, and a plurality of trained teacher identification networks, and identify a data type of each customer service data, and a data type corresponding to each teacher identification network;

the calculating module 320 is configured to perform data marking processing on each customer service data of the data type based on the identification teacher network of the data type, obtain pseudo tag information of each customer service data, and calculate a noise value of each customer service data based on the pseudo tag information of each customer service data and the real tag information of each customer service data;

a screening module 330, configured to screen customer service data corresponding to a noise value lower than a noise threshold, as target customer service data of the data type, and generate a transition probability matrix of the data type based on pseudo tag information of each target customer service data and transition probability information between the pseudo tag information and real tag information of each target customer service data;

The training module 340 is configured to train, through a knowledge distillation loss function, classification parameters of an initial recognition student model corresponding to the data type based on pseudo tag information of each target customer service data, a transition probability matrix of the data type, a recognition teacher network of the data type, and each target customer service data, so as to obtain a recognition student model corresponding to the data type;

the construction module 350 is configured to obtain a scoring policy for each data type, and construct a customer service scoring model based on the identification model corresponding to each data type and the scoring policy for each data type.

Optionally, the acquiring module 310 is specifically configured to:

extracting data identification information of each customer service data, and identifying the data type of each customer service data based on the data identification information of each customer service data;

acquiring training sample data of each recognition teacher network and data identification information of each training sample data, and recognizing the data type of each training sample data based on the data identification information of each training sample data; the training process of the teacher network is that the local client performs parameter training processing on each initial teacher network based on training sample data of each teacher network to obtain each teacher network;

And taking the data type of each training sample data as the data type of the recognition teacher network trained by each training sample data.

Optionally, the computing module 320 is specifically configured to:

based on the data type recognition teacher network, predicting the predicted data category of each customer service data respectively, and taking the predicted data category of each customer service data as pseudo tag information of each customer service data;

and respectively calculating the noise value of each customer service data through a label-free data noise value algorithm based on each customer service data and the pseudo label information of each customer service data.

Optionally, the screening module 330 is specifically configured to:

identifying the data category of the pseudo tag information of each target customer service data and the data category corresponding to the real tag information of each target customer service data;

based on the target customer service data to which each piece of pseudo tag information belongs, the target customer service data to which each piece of real tag information belongs, the data category corresponding to each piece of pseudo tag information, and the data category corresponding to each piece of real tag information, respectively calculating the conversion probability of each piece of real tag information into the pseudo tag information through a probability calculation formula;

And constructing a transition probability matrix of the data type based on the transition probability of each real tag information to a pseudo tag information, each pseudo tag information and each real tag information.

Optionally, the training module 340 is specifically configured to:

based on the transition probability matrix of the data type, adjusting an initial knowledge distillation loss function to obtain a knowledge distillation loss function corresponding to the data type;

based on the pseudo tag information of each target customer service data, the identification teacher network of the data type and each target customer service data, training the classification parameters of the initial identification student model corresponding to the data type through the knowledge distillation loss function of the data type to obtain the identification student model corresponding to the data type.

Optionally, the construction module 350 is specifically configured to:

the method comprises the steps of obtaining grading information of real tag information of each data type and grading weight values of each data type, and respectively adding the grading information of the real tag information of each data type to an identification student model corresponding to each data type to obtain customer service grading student models corresponding to each data type;

And determining weight parameters of customer service scoring student models corresponding to the data types based on the scoring weight values of the data types, and summarizing the customer service scoring student models and the weight parameters of the customer service scoring student models to obtain the customer service scoring models.

The modules in the customer service scoring model building device can be all or partially realized by software, hardware and a combination thereof. The above modules may be embedded in hardware or may be independent of a processor in the computer device, or may be stored in software in a memory in the computer device, so that the processor may call and execute operations corresponding to the above modules.

In one embodiment, a computer device is provided, which may be a terminal, and the internal structure of which may be as shown in fig. 4. The computer device includes a processor, a memory, a communication interface, a display screen, and an input device connected by a system bus. Wherein the processor of the computer device is configured to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The communication interface of the computer device is used for carrying out wired or wireless communication with an external terminal, and the wireless mode can be realized through WIFI, a mobile cellular network, NFC (near field communication) or other technologies. The computer program when executed by a processor implements a method of building a customer service scoring model. The display screen of the computer equipment can be a liquid crystal display screen or an electronic ink display screen, and the input device of the computer equipment can be a touch layer covered on the display screen, can also be keys, a track ball or a touch pad arranged on the shell of the computer equipment, and can also be an external keyboard, a touch pad or a mouse and the like.

Those skilled in the art will appreciate that the structures shown in FIG. 4 are block diagrams only and do not constitute a limitation of the computer device on which the present aspects apply, and that a particular computer device may include more or less components than those shown, or may combine some of the components, or have a different arrangement of components.

In an embodiment, a computer device is provided comprising a memory and a processor, the memory having stored therein a computer program, the processor implementing the steps of the method of any of the first aspects when the computer program is executed.

In an embodiment, a computer-readable storage medium is provided, on which a computer program is stored which, when executed by a processor, implements the steps of the method of any of the first aspects.

In an embodiment, a computer program product is provided, comprising a computer program which, when executed by a processor, implements the steps of the method of any of the first aspects.

It should be noted that, user information (including but not limited to user equipment information, user personal information, etc.) and data (including but not limited to data for analysis, stored data, presented data, etc.) referred to in the present application are information and data authorized by the user or sufficiently authorized by each party.

Those skilled in the art will appreciate that implementing all or part of the above described methods may be accomplished by way of a computer program stored on a non-transitory computer readable storage medium, which when executed, may comprise the steps of the embodiments of the methods described above. Any reference to memory, database, or other medium used in the various embodiments provided herein may include at least one of non-volatile and volatile memory. The nonvolatile Memory may include Read-Only Memory (ROM), magnetic tape, floppy disk, flash Memory, optical Memory, high density embedded nonvolatile Memory, resistive random access Memory (ReRAM), magnetic random access Memory (Magnetoresistive Random Access Memory, MRAM), ferroelectric Memory (Ferroelectric Random Access Memory, FRAM), phase change Memory (Phase Change Memory, PCM), graphene Memory, and the like. Volatile memory can include random access memory (Random Access Memory, RAM) or external cache memory, and the like. By way of illustration, and not limitation, RAM can be in the form of a variety of forms, such as static random access memory (Static Random Access Memory, SRAM) or dynamic random access memory (Dynamic Random Access Memory, DRAM), and the like. The databases referred to in the various embodiments provided herein may include at least one of relational databases and non-relational databases. The non-relational database may include, but is not limited to, a blockchain-based distributed database, and the like. The processors referred to in the embodiments provided herein may be general purpose processors, central processing units, graphics processors, digital signal processors, programmable logic units, quantum computing-based data processing logic units, etc., without being limited thereto.

The technical features of the above embodiments may be arbitrarily combined, and all possible combinations of the technical features in the above embodiments are not described for brevity of description, however, as long as there is no contradiction between the combinations of the technical features, they should be considered as the scope of the description.

The above examples only represent a few embodiments of the present application, which are described in more detail and are not to be construed as limiting the scope of the present application. It should be noted that it would be apparent to those skilled in the art that various modifications and improvements could be made without departing from the spirit of the present application, which would be within the scope of the present application. Accordingly, the scope of protection of the present application shall be subject to the appended claims.

Claims (10)

1. The method for constructing the customer service scoring model is characterized by comprising the following steps of:

acquiring a plurality of unlabeled customer service data, real label information of each customer service data and a plurality of trained identification teacher networks, and identifying the data type of each customer service data and the data type corresponding to each identification teacher network;

for each data type, based on an identification teacher network of the data type, respectively carrying out data marking processing on each customer service data of the data type to obtain pseudo tag information of each customer service data, and calculating a noise value of each customer service data based on the pseudo tag information of each customer service data and the real tag information of each customer service data;

Screening customer service data corresponding to a noise value lower than a noise threshold value, serving as target customer service data of the data type, and generating a transition probability matrix of the data type based on pseudo tag information of each target customer service data and transition probability information between real tag information of each target customer service data;

training classification parameters of an initial recognition student model corresponding to the data type through a knowledge distillation loss function based on pseudo tag information of each target customer service data, a conversion probability matrix of the data type, a recognition teacher network of the data type and each target customer service data to obtain the recognition student model corresponding to the data type;

and obtaining a scoring strategy of each data type, and constructing a customer service scoring model based on the identification model corresponding to each data type and the scoring strategy of each data type.

2. The method of claim 1, wherein the identifying the data type of each customer service data and the data type corresponding to each identified teacher network comprises:

extracting data identification information of each customer service data, and identifying the data type of each customer service data based on the data identification information of each customer service data;

Acquiring training sample data of each recognition teacher network and data identification information of each training sample data, and recognizing the data type of each training sample data based on the data identification information of each training sample data; the training process of the teacher network is that the local client performs parameter training processing on each initial teacher network based on training sample data of each teacher network to obtain each teacher network;

and taking the data type of each training sample data as the data type of the recognition teacher network trained by each training sample data.

3. The method according to claim 2, wherein the identifying teacher network based on the data type performs data marking processing on each customer service data of the data type to obtain pseudo tag information of each customer service data, and calculates a noise value of each customer service data based on the pseudo tag information of each customer service data and the real tag information of each customer service data, including:

based on the data type recognition teacher network, predicting the predicted data category of each customer service data respectively, and taking the predicted data category of each customer service data as pseudo tag information of each customer service data;

And respectively calculating the noise value of each customer service data through a label-free data noise value algorithm based on each customer service data and the pseudo label information of each customer service data.

4. The method of claim 1, wherein generating the transition probability matrix of the data type based on transition probability information between pseudo tag information of each of the target customer service data and real tag information of each of the target customer service data comprises:

identifying the data category of the pseudo tag information of each target customer service data and the data category corresponding to the real tag information of each target customer service data;

based on the target customer service data to which each piece of pseudo tag information belongs, the target customer service data to which each piece of real tag information belongs, the data category corresponding to each piece of pseudo tag information, and the data category corresponding to each piece of real tag information, respectively calculating the conversion probability of each piece of real tag information into the pseudo tag information through a probability calculation formula;

and constructing a transition probability matrix of the data type based on the transition probability of each real tag information to a pseudo tag information, each pseudo tag information and each real tag information.

5. The method according to claim 1, wherein training the classification parameters of the initial recognition student model corresponding to the data type by the knowledge distillation loss function based on the pseudo tag information of each target customer service data, the transition probability matrix of the data type, the recognition teacher network of the data type, and each target customer service data to obtain the recognition student model corresponding to the data type comprises:

based on the transition probability matrix of the data type, adjusting an initial knowledge distillation loss function to obtain a knowledge distillation loss function corresponding to the data type;

based on the pseudo tag information of each target customer service data, the identification teacher network of the data type and each target customer service data, training the classification parameters of the initial identification student model corresponding to the data type through the knowledge distillation loss function of the data type to obtain the identification student model corresponding to the data type.

6. The method of claim 1, wherein the constructing a customer service scoring model based on the identification model corresponding to each data type and the scoring policy for each data type comprises:

The method comprises the steps of obtaining grading information of real tag information of each data type and grading weight values of each data type, and respectively adding the grading information of the real tag information of each data type to an identification student model corresponding to each data type to obtain customer service grading student models corresponding to each data type;

and determining weight parameters of customer service scoring student models corresponding to the data types based on the scoring weight values of the data types, and summarizing the customer service scoring student models and the weight parameters of the customer service scoring student models to obtain the customer service scoring models.

7. A customer service scoring model construction device, the device comprising:

the system comprises an acquisition module, a processing module and a processing module, wherein the acquisition module is used for acquiring a plurality of unlabeled customer service data, real label information of each customer service data and a plurality of trained recognition teacher networks, and recognizing the data type of each customer service data and the data type corresponding to each recognition teacher network;

the computing module is used for respectively carrying out data marking processing on each customer service data of the data type according to the identification teacher network of the data type to obtain pseudo tag information of each customer service data, and computing the noise value of each customer service data according to the pseudo tag information of each customer service data and the real tag information of each customer service data;

The screening module is used for screening customer service data corresponding to a noise value lower than a noise threshold value, serving as target customer service data of the data type, and generating a transition probability matrix of the data type based on pseudo tag information of each target customer service data and transition probability information between the pseudo tag information and real tag information of each target customer service data;

the training module is used for training the classification parameters of the initial recognition student model corresponding to the data type through a knowledge distillation loss function based on the pseudo tag information of each target customer service data, the conversion probability matrix of the data type, the recognition teacher network of the data type and each target customer service data to obtain the recognition student model corresponding to the data type;

the construction module is used for acquiring the scoring strategies of the data types and constructing a customer service scoring model based on the identification model corresponding to the data types and the scoring strategies of the data types.

8. A computer device comprising a memory and a processor, the memory storing a computer program, characterized in that the processor implements the steps of the method of any of claims 1 to 6 when the computer program is executed.

9. A computer readable storage medium, on which a computer program is stored, characterized in that the computer program, when being executed by a processor, implements the steps of the method of any of claims 1 to 6.

10. A computer program product comprising a computer program, characterized in that the computer program, when being executed by a processor, implements the steps of the method of any of claims 1 to 6.